Hybrid analysis of photovoltaic energy data containing excessive structural zeros

Abstract

Abstract Photovoltaic (PV) power generation plays a critical role in the global transition toward sustainable energy systems. However, accurate PV power forecasting remains challenging due to the non-stationary nature of PV power time series and the presence of structural zeros. Most existing studies rely on parametric models or treat zero inflation as a secondary issue. In this study, a hybrid hurdle modeling framework is implemented to explicitly account for both structural zeros and continuous positive power values through a two-part structure. The zero component is modeled using Bayesian logistic regression (BLR), random forest classifier (RFC), and support vector classifier (SVC). The positive values are analyzed using the parametric models Gamma regression (GR), Log-normal regression (LNR), and Weibull regression (WR) and the non-parametric machine learning (ML) techniques random forest regression (RFR), extreme gradient boosting (XGBoost), and support vector regression (SVR). The proposed framework is validated using the target variable active power (AP, kW) based on a real-world data from a 110 kWe (129.6 kWp) PV plant located in Çaycuma, Zonguldak, Türkiye. Four metrics were used to compare the predictive performances of the models: mean squared error (MSE), mean absolute (scaled) error (MAE and MASE), and the coefficient of determination ( $$\:{R}^{2}$$ ). Among the 18 distinct hurdle models, the lowest error values and the highest $$\:{R}^{2}$$ were obtained when the positive component was modeled by either RFR, such as BLR-RFR (MSE = 67.425 kW, MAE = 3.974 kW, MASE = 0.615, $$\:{R}^{2}$$ = 0.899) or XGBoost, such as RFC-XGBoost (MSE = 69.252 kW, MAE = 4.075 kW, MASE = 0.630, $$\:{R}^{2}$$ = 0.898). The findings indicate that hybrid modeling approaches where the zero component is modeled using BLR, RFC, or SVC, and the positive component is modeled with either RFR or XGBoost provide a reliable framework for PV power forecasting, when predicting continuous data characterized by excessive structural zeros.